Prediction and evaluation of photovoltaic-thermal energy systems production using artificial neural network and experimental dataset

• Main Authors: Jabar H. Yousif, Hussein A. Kazem
• Format: Article
• Language: English
• Published: Elsevier 2021-10-01
• Series: Case Studies in Thermal Engineering
• Subjects: Solar energy; Photovoltaic performance; Hybrid PV/T; Energy prediction; ANN
• Online Access: http://www.sciencedirect.com/science/article/pii/S2214157X21004603

Photovoltaic/thermal (PV/T) systems combine two collectors, which increase efficiency, reduce cost and space, and produce electricity and heat, simultaneously. Many factors affect PV/T current, voltage, power, efficiency, and heat energy production. For example, the location of the PV system, ambient temperature, irradiance, humidity, dust, and many other factors. Also, different modelling techniques are used to evaluate PV/T efficiency, for example, analytical, regression, numerical, artificial neural network (ANN). The current work aims to predict and assess a PV/T system using ANN models based on an experimental dataset in Oman. The PV/T system with weather station and data acquisition was installed in Sohar, Oman. The weather and electrical data has been recorded. A novel mathematical and ANN model for examining the performance of PV/T systems has been developed. The experimental results show improvement in PVT power production (68.6132 W) compared to the conventional PV (66.7827 W). The results demonstrate that the three proposed models (MLP, SOFM, and SVM) achieved excellent MSE results for generating the current values of the PV system (0.00043, 0.00030, 0.00041) and PV/T system (0.00719, 0.00683, 0.00763), respectively. Also, the proposed models delivered excellent MSE results for simulating the power values of the PV system (0.04457, 0.05006, 0.13816) and PV/T system (0.04457, 0.05006, 0.13816), respectively. The proposed models result validated with experimental data using descriptive statistics and Evaluation Metrics. Finally, the proposed neural models can generate future figures for any needed period that accurately fit the actual datasets.